How the system is validated and governed

Implementation details, parameters and instrument selection are deliberately not published. That is a decision, not an omission.

What is published is the standard the system is measured against — which is the part worth reading anyway. A method survives scrutiny; a rule set only survives secrecy.

Parameters are chosen on an earlier window and judged only on a later window the selection never saw. The pair is then rolled forward and the process repeats, so that every result reported anywhere on this site is out-of-sample.

This produced 14 out-of-sample windows spanning January 2018 to May 2026.

Fees on both sides of a position, slippage, and funding or financing charges are all deducted before a result counts as a result. A gross number is not evidence of anything.

Monte Carlo across 1,000 trade-order reshuffles, to separate the shape of the outcome from the order it happened to arrive in.

Sensitivity checking confirms that performance does not collapse under small perturbation — a configuration must remain acceptable in the neighbourhood around it, not only at the exact point that tested best.

Fixed in writing before results were seen. Pre-commitment is the entire point of the section:

• Profit factor at least 1.3
• Average trade at least twice its own cost
• Max drawdown no worse than 20%
• Monte Carlo p95 drawdown no worse than 25%
• Sharpe at least 0.8, or MAR at least 0.5
• At least 100 out-of-sample trades
• A stress test at triple slippage and double funding still above 1.1

Configurations failing any single criterion were rejected regardless of how good the rest of the profile looked.

Changes are proposed within bounded limits, reviewed before they take effect, and never applied automatically in a way that increases risk. Anything that could loosen a constraint requires a human to act; nothing loosens itself.

• Backtests are not live trading.
• Simulated fills do not model real market impact.
• The test period is dominated by a small number of large market moves, and results depend on comparable conditions recurring.
• The traded universe is narrow.
• Walk-forward reduces but cannot eliminate selection bias.